[PATCH] epoll: try to be a _bit_ better about file lifetimes

[Linus Torvalds]

On Fri, 3 May 2024 at 16:23, Kees Cook <keescook at chromium.org> wrote:
>
> static bool __must_check get_dma_buf_unless_doomed(struct dma_buf *dmabuf)
> {
>         return atomic_long_inc_not_zero(&dmabuf->file->f_count) != 0L;
> }
>
> If we end up adding epi_fget(), we'll have 2 cases of using
> "atomic_long_inc_not_zero" for f_count. Do we need some kind of blessed
> helper to live in file.h or something, with appropriate comments?

I wonder if we could try to abstract this out a bit more.

These games with non-ref-counted file structures *feel* a bit like the
games we play with non-ref-counted (aka "stashed") 'struct dentry'
that got fairly recently cleaned up with path_from_stashed() when both
nsfs and pidfs started doing the same thing.

I'm not loving the TTM use of this thing, but at least the locking and
logic feels a lot more straightforward (ie the
atomic_long_inc_not_zero() here is clealy under the 'prime->mutex'
lock

IOW, the tty use looks correct to me, and it has fairly simple locking
and is just catching the the race between 'fput()' decrementing the
refcount and and 'file->f_op->release()' doing the actual release.

You are right that it's similar to the epoll thing in that sense, it
just looks a _lot_ more straightforward to me (and, unlike epoll,
doesn't look actively buggy right now).

Could we abstract out this kind of "stashed file pointer" so that we'd
have a *common* form for this? Not just the inc_not_zero part, but the
locking rule too?

              Linus